PhD Researcher, Mechanical Engineering
University of Washington, AI Institute in Dynamic Systems
Advised by Prof. Stefania Fresca
I build reliable models and controllers for physical systems by auditing neural surrogates against high-fidelity simulations.
Sim2Science workshop at NeurIPS 2026 · Accepted
Audits surrogate action gradients against high-fidelity finite differences, extracts a compact feedback law, and anchors policy optimization to that law before validating controllers in the original solver.
Journal manuscript: In preparation for submission to TMLR.
Learning for Dynamics and Control Conference (L4DC)
Studies when temporal history can replace spatial context in forecasting from sparse or noisy sensors, using exact Gaussian conditioning and learned models of linear fields, Burgers flow, and the Kuramoto–Sivashinsky equation.
Submitted to L4DC.
JGR: Machine Learning and Computation, 2(2)
Compares four architectures for next day wildfire segmentation. SHAP, Grad-CAM, and Integrated Gradients characterize architecture-dependent feature attribution.
Audits surrogate action gradients against high-fidelity finite differences, extracts a compact feedback law, and anchors policy optimization to that law before validating controllers in the original solver.
Studies when temporal history can replace spatial context in forecasting from sparse or noisy sensors, using exact Gaussian conditioning and learned models of linear fields, Burgers flow, and the Kuramoto–Sivashinsky equation.
Comparison of CNN and transformer architectures under class imbalance, with global and spatial feature attribution.
Contributions to the FISICA project. Further details are limited due to confidentiality.
University of Washington, SCiDaMa Group
Develop neural surrogates and reinforcement learning methods for high-dimensional flow control. Audit learned action sensitivities and validate controllers in direct numerical simulation.
University of Cambridge
Studied nonlocal state inference in neural PDE surrogates from partial and spatially coarsened observations.
Imperial College London, Data Science Institute
Compared CNN and transformer architectures for wildfire segmentation and analyzed feature attribution with SHAP, Grad-CAM, and Integrated Gradients.
J.P. Morgan, Quantitative Solutions
Built transformer research pipelines for financial-news sentiment analysis and stock-return forecasting.
European Bioinformatics Institute, Goldman Group
Developed synthetic-data simulations and CNN and transformer models for biological sequence classification.
University of Washington, AI Institute in Dynamic Systems
Scientific machine learning, fluid mechanics, reinforcement learning, and optimal control.1st Place
AIGC Hackbridge UK Hackathon
2nd Place
Citadel European Datathon
Gold Medal
iGEM Synthetic Biology Competition
1st Place
AIGC Hackbridge UK Hackathon
Top 1%
IMC Quantitative Trading Competition